"AI Hacker" is coming, how can Agentic AI become the new guardian?

Author: Pumping Geek

01 The Rise of AI: The Security Shadow War Under the Double-Edged Sword of Technology

With the rapid development of AI technology, the threats faced by cybersecurity are becoming increasingly complex. Attack methods are not only more efficient and covert, but have also given rise to a new form of "AI hackers," leading to various new types of cybersecurity crises.

First of all, generative AI is reshaping the "precision" of online scams.

In simple terms, it involves the intelligent automation of traditional phishing attacks. For example, in more precise scenarios, attackers use publicly available social data to train AI models to generate personalized phishing emails in bulk, mimicking the writing style or language habits of specific users, implementing "customized" scams that bypass traditional spam filters and significantly increase the success rate of attacks.

Next is the most well-known deepfake and identity impersonation. Before AI technology matured, traditional "face-swapping fraud attacks," also known as BEC fraud, which stands for "Business Email Compromise," specifically involved attackers disguising the email sender as your leader, colleague, or business partner in order to defraud business information or money, or to obtain other important materials.

Nowadays, "face-changing" has really happened. AI-generated face-swapping and voice-changing technologies can forge the identities of public figures or friends and family, used for fraud, public opinion manipulation, and even political interference. Just two months ago, the financial director of a company in Shanghai received a video conferencing invitation from the "chairman". The other party used AI face-swapping and voice imitation to claim that an urgent payment of "overseas cooperation deposit" was needed. Following the instructions, the director transferred 3.8 million to the designated account, later realizing it was an overseas fraud group using deep forgery technology to commit the crime.

The third is automated attacks and vulnerability exploitation. The advancement of AI technology has led to the evolution of numerous scenarios towards intelligence and automation, and cyber attacks are no exception. Attackers can use AI to automatically scan for system vulnerabilities, generate dynamic attack codes, and carry out indiscriminate rapid attacks on targets. For example, AI-driven "zero-day attacks" can immediately write and execute malicious programs upon discovering vulnerabilities, making it difficult for traditional defense systems to respond in real-time.

During this year's Spring Festival, the DeepSeek official website suffered a massive DDoS attack of 3.2Tbps. Hackers simultaneously infiltrated through the API to inject adversarial samples, altering model weights and causing core services to be paralyzed for 48 hours, resulting in direct economic losses exceeding tens of millions of dollars. Subsequent investigations revealed traces of long-term infiltration by the US NSA.

Data poisoning and model vulnerabilities are also a new threat. Attackers can inject false information into AI training data (i.e., data poisoning) or exploit flaws in the model itself to induce AI to output incorrect results—this poses direct security threats to critical areas and could even trigger a chain of catastrophic consequences, such as an autonomous driving system misjudging "no entry" as a "speed limit sign" due to adversarial samples, or medical AI misclassifying benign tumors as malignant.

02 AI still needs AI governance

In the face of new AI-driven cybersecurity threats, traditional protection models have proven inadequate. So, what countermeasures do we have?

It is not hard to see that the current industry consensus points to "using AI to combat AI"—this is not only an upgrade of technical means but also a transformation of the security paradigm.

Existing attempts can be roughly divided into three categories: security protection technologies for AI models, industry-level defense applications, and more macro-level government and international collaboration.

The key to AI model security protection technology lies in the intrinsic security reinforcement of the model.

Taking the "jailbreak" vulnerability of large language models (LLMs) as an example, their security protection mechanisms often fail due to generic jailbreak prompt strategies—attackers systematically bypass the model's built-in protective layers, inducing the AI to generate violent, discriminatory, or illegal content. To prevent the "jailbreak" of LLMs, various model companies have made attempts, such as Anthropic, which released a "Constitution Classifier" in February this year.

Here, the term "Constitution" refers to the inviolable natural language rules, serving as a safeguard trained on synthetic data. By specifying allowed and restricted content, it monitors input and output in real time. Under baseline condition testing, its Claude3.5 model, protected by classifiers, increased the success prevention rate of advanced jailbreak attempts from 14% to 95%, significantly reducing the AI's "jailbreak" risk.

In addition to model-based and more general defense measures, industry-level defense applications are also worth paying attention to, as their scenario-based protection in vertical fields is becoming a key breakthrough point: the financial industry builds anti-fraud barriers through AI risk control models and multimodal data analysis, the open-source ecosystem achieves rapid response to zero-day threats by leveraging intelligent vulnerability hunting technology, while the protection of sensitive information in enterprises relies on an AI-driven dynamic control system.

For example, the solution showcased by Cisco at the Singapore International Network Week can intercept sensitive data query requests submitted by employees to ChatGPT in real-time and automatically generate compliance audit reports to optimize the management loop.

At the macro level, government and international cross-regional collaboration is also accelerating. The Cyber Security Agency of Singapore has released the "Guidelines for the Security of Artificial Intelligence Systems," which constrains the misuse of generative AI through mandatory localization deployment and data encryption mechanisms, particularly establishing protective standards for identifying AI-generated identities in phishing attacks; the United States, the United Kingdom, and Canada have simultaneously launched the "AI Network Agent Program," focusing on the research and development of trustworthy systems and real-time assessments of APT attacks, enhancing collective defense capabilities through a unified security certification system.

So, what methods can maximize the use of AI to tackle the cybersecurity challenges of the AI era?

"The future requires AI secure intelligent hubs and the construction of new systems around the hubs." At the 2nd Wuhan Cybersecurity Innovation Forum, Zhang Fu, founder of Qingteng Cloud Security, emphasized that using AI to combat AI is the core of the future cybersecurity defense system. "Within 3 years, AI will disrupt the existing security industry and all 2B industries. Products will be restructured to achieve unprecedented efficiency and capability improvements. Future products are designed for AI, not for humans."

Among various solutions, the model of Security Copilot clearly provides a good demonstration of "future products are for AI": A year ago, Microsoft launched the intelligent Microsoft Security Copilot co-pilot to help security teams quickly and accurately detect, investigate, and respond to security incidents; a month ago, it released an AI agent that automatically assists in key areas such as phishing attacks, data security, and identity management.

Microsoft has added six self-developed AI agents to expand the functionality of Security Copilot. Three of them assist cybersecurity personnel in filtering alerts: the phishing classification agent reviews phishing alerts and filters false positives; the other two analyze Purview notifications to detect unauthorized use of business data by employees.

The Conditional Access Optimization Agent collaborates with Microsoft Entra to identify unsafe user access rules and generate one-click remediation plans for administrators to execute. The Vulnerability Remediation Agent integrates with the device management tool Intune to help quickly locate vulnerable endpoints and apply operating system patches. The Threat Intelligence Briefing Agent generates cybersecurity threat reports that may threaten the organization's systems.

03 Wu Xiang: The Escort of L4 Advanced Intelligent Agents

Coincidentally, in China, in order to achieve true "autonomous driving" level safety protection, Qingteng Cloud Security has launched a full-stack security intelligence agent called "Wuxiang". As the world's first safety AI product to transition from "assisted AI" to "autonomous agent" (Autopilot), its core breakthrough lies in overturning the traditional tool's "passive response" model, making it autonomous, automatic, and intelligent.

By integrating machine learning, knowledge graphs, and automated decision-making technologies, "Wuxiang" can independently complete the entire process loop from threat detection, impact assessment to response management, achieving true autonomous decision-making and goal-driven actions. Its "Agentic AI architecture" design simulates the collaborative logic of human security teams: using the "brain" to integrate the cybersecurity knowledge base to support planning capabilities, the "eyes" to finely perceive the dynamics of the network environment, and the "hands and feet" to flexibly utilize a diverse security toolchain, forming an efficient assessment network for information sharing through multi-agent collaboration, with division of labor and information sharing.

In terms of technical implementation, "No Phase" adopts the "ReAct Mode" (Act-Observe-Think-Act cycle) and the "Plan AI + Action AI dual-engine architecture" to ensure dynamic correction capabilities in complex tasks. When there is an abnormality in tool invocation, the system can autonomously switch to a backup plan instead of interrupting the process. For example, in APT attack analysis, Plan AI acts as the "organizer" to break down task objectives, while Action AI functions as the "investigation expert" to execute log parsing and threat modeling, with both advancing in parallel based on a real-time shared knowledge graph.

In terms of functional modules, "Wuxiang" has built a complete autonomous decision-making ecosystem: the intelligent body simulates the reflective iterative thinking of a security analyst, dynamically optimizing the decision-making path; tool integration integrates host security log queries, network threat intelligence retrieval, and LLM-driven malware analysis; environmental awareness captures host assets and network information in real time; knowledge graphs dynamically store entity associations to assist decision-making; multi-agent collaboration executes tasks in parallel through task decomposition and information sharing.

Currently, "Wuxiang" performs the best in three core application scenarios: alarm assessment, traceability analysis, and outputting security reports.

In traditional security operations, the verification of the authenticity of massive alerts is time-consuming and labor-intensive. Taking a local privilege escalation alert as an example: the non-entity alert analysis intelligent agent automatically parses threat characteristics, invokes process privilege analysis, parent process tracing, program signature verification, and other toolchains, ultimately determining it as a false positive—without any human intervention throughout the process. In the existing alert testing of Qingteng, the system has achieved a 100% alert coverage rate and a 99.99% analysis accuracy rate, significantly reducing manual workload by over 95%.

In the face of real threats such as Webshell attacks, the intelligent agent confirms the validity of attacks in seconds through cross-dimensional correlation such as code feature extraction and file permission analysis. Traditional in-depth tracing that requires multi-department collaboration and takes several days (such as upload propagation path restoration and lateral impact assessment) is now automatically connected by the system using data streams from host logs, network traffic, and behavioral baselines, generating a complete attack chain report and compressing the response cycle from "days" to "minutes."

"Our core is to reverse the relationship of cooperation between AI and humans, allowing AI to cooperate as if it were a person, achieving a leap from L2 to L4, that is, from assisted driving to high-level autonomous driving." said Hu Jun, co-founder of Qingteng and vice president of products. "As AI can adapt to more scenarios and the success rate of decision-making increases, it will gradually be able to take on more responsibilities, thus changing the division of responsibilities between humans and AI."

In the scenario of traceability analysis, the first step is the triggering of a Webshell alert, which initiates a multi-agent security team collaboration driven by "no-phase AI" for traceability: the "judgment expert" locates the one.jsp file based on the alert, generating parallel tasks such as file content analysis, author traceability, directory investigation, and process tracking. The "security investigator" agent utilizes file log tools to quickly identify the java (12606) process as the source of the write operation, and this process along with the associated host 10.108.108.23 (identified through access logs showing high-frequency interactions) are subsequently included in the investigation.

The intelligent agent dynamically expands clues through a threat graph, drilling down layer by layer from a single file to processes and hosts, assessing the summarized task results from experts to comprehensively determine risks. This process compresses what would take human investigators hours to days into just a few minutes, restoring the entire attack chain with accuracy that surpasses that of human senior security experts, tracking lateral movement paths without any blind spots. Red team assessments also indicate that it is difficult to evade its comprehensive investigation.

"Large models outperform humans because they can thoroughly investigate every nook and cranny, rather than relying on experience to rule out low-probability scenarios," Hu Jun explained. "This is equivalent to having better breadth and depth."

After investigating complex attack scenarios, organizing alerts and investigation clues and generating reports is often time-consuming and labor-intensive. AI can achieve one-click summarization, clearly presenting the attack process in the form of a visual timeline, coherently displaying key nodes like a movie—the system will automatically sort key evidence to generate key frames of the attack chain, combining environmental context information, ultimately generating a dynamic attack chain diagram that presents the entire attack trajectory in an intuitive and three-dimensional manner.

04 Conclusion

It is clear that the development of AI technology brings dual challenges to cybersecurity.

On one hand, attackers use AI to automate, personalize, and conceal attacks; on the other hand, defenders need to accelerate technological innovation and enhance detection and response capabilities through AI. In the future, the AI technology competition between attackers and defenders will determine the overall situation of cybersecurity, and the improvement of security intelligence will be key to balancing risk and development.

The secure intelligent agent "Wuxiang" has brought new changes at both the security architecture and cognitive levels.

"No Form" essentially changes the way AI is used. Its breakthrough lies in the fusion of multidimensional data perception, protective strategy generation, and decision-making interpretability into an organic whole—transforming the past model of using AI as a tool into empowering AI to work independently and automatically.

By analyzing heterogeneous data such as logs, texts, and traffic through correlation analysis, the system can capture traces of APT activities before the attacker builds a complete attack chain. More importantly, the visualization of its decision-making process and reasoning explanation transforms traditional tools' black box alerts, which only indicate what is happening without understanding why, into history—security teams can not only see threats but also understand the evolution logic of those threats.

The essence of this innovation is a paradigm shift in security thinking from "fixing the barn after the sheep are lost" to "preparing for a rainy day," which is a redefinition of the rules of offense and defense.

"Wuxiang" is like a hunter with digital intuition: by real-time modeling of memory operations and other microscopic behavioral features, it can extract latent custom trojans from massive noise; the dynamic attack surface management engine continuously assesses asset risk weights, ensuring that protective resources are accurately directed at critical systems; while the intelligent digestion mechanism of threat intelligence converts an average of tens of thousands of alerts per day into actionable defense instructions, even predicting the evolutionary direction of attack variants—while traditional solutions are still struggling to cope with existing intrusions, "Wuxiang" has already anticipated and blocked the next move of the attackers.

The birth of the "AI Intelligent Central System (Advanced Security Intelligence Agent)" will completely reshape the landscape of cybersecurity. And all we need to do is seize this opportunity completely," said Zhang Fu.